Development of predictive models for sustainable concrete via genetic programming-based algorithms

Abstract

Waste foundry sand (WFS), a by-product of the casting industry, is a potential material that may be employed as a substitute for fine aggregate in concrete. In the present study, gene expression programming (GEP) and multi-expression programming (MEP) are used to generate predictive models for the split tensile strength (STS) and elastic modulus (E) of waste foundry sand concrete (WFSC). Therefore, a comprehensive database was collected that contains 146 and 242 values of E and STS, respectively. Seven different variables were chosen as input for the development of the ML-based models. The reliability and accuracy of the proposed model were evaluated by using various statistical indicators. Given the performance assessment, both GEP and MEP accurately predict the E with a correlation of 0.994 and 0.996, respectively. However, GEP performance was much superior in predicting STS (R = 0.987) as compared to the MEP model (R = 0.892). The integrated statistical performance (ρ, OF) of both models approaches zero, indicating the excellent performance and generalization potential of the developed models. For the interpretation of machine learning (ML) models, Shapley additive explanation (SHAP) was used to know about the input variables' importance and influence on the output parameter. The SHAP analysis revealed that a higher ratio of FA/TA results in the enhancement of the elastic modulus, whereas CA/C higher ratio is favorably influencing the split tensile strength up to some extent, however, this trend changes when the ratio is further increased. These soft computing prediction techniques can incentivize the use of WFS in sustainable concrete, reducing waste disposal and promoting environment-friendly construction. Furthermore, it is recommended that the findings of this study be validated with more extensive data sets and that other ML techniques be investigated.